{"metadata":{"kernelspec":{"language":"python","display_name":"Python 3","name":"python3"},"language_info":{"name":"python","version":"3.10.12","mimetype":"text/x-python","codemirror_mode":{"name":"ipython","version":3},"pygments_lexer":"ipython3","nbconvert_exporter":"python","file_extension":".py"},"kaggle":{"accelerator":"none","dataSources":[{"sourceId":21669,"databundleVersionId":1692278,"sourceType":"competition"}],"dockerImageVersionId":30839,"isInternetEnabled":true,"language":"python","sourceType":"notebook","isGpuEnabled":false}},"nbformat_minor":4,"nbformat":4,"cells":[{"cell_type":"code","source":"!pip install efficientnet_pytorch","metadata":{"_uuid":"8f2839f25d086af736a60e9eeb907d3b93b6e0e5","_cell_guid":"b1076dfc-b9ad-4769-8c92-a6c4dae69d19","trusted":true,"execution":{"iopub.status.busy":"2025-01-28T23:08:36.022848Z","iopub.execute_input":"2025-01-28T23:08:36.023193Z","iopub.status.idle":"2025-01-28T23:08:44.246913Z","shell.execute_reply.started":"2025-01-28T23:08:36.023157Z","shell.execute_reply":"2025-01-28T23:08:44.245754Z"}},"outputs":[],"execution_count":null},{"cell_type":"code","source":"import torch\nimport torch.nn as nn\nimport numpy as np\nimport random\nimport copy\nimport warnings\nimport torch\nimport librosa\nimport csv\nimport os\nimport pandas as pd\n\nfrom skimage.transform import resize\nfrom skimage.filters import gaussian\nfrom skimage.color import rgb2gray\nfrom skimage import exposure, util\nfrom efficientnet_pytorch import EfficientNet\nfrom torch.utils.data import Dataset, DataLoader\nfrom tqdm import tqdm\nfrom sklearn.model_selection import KFold","metadata":{"trusted":true},"outputs":[],"execution_count":null},{"cell_type":"code","source":"class AudioData(Dataset):\n    def __init__(self, X, y, data_type, audio_data, fmin, fmax, length):\n        self.data = []\n        self.labels = []\n        self.augs = [addNoisy, contrast_stretching, randomGaussian, randomGamma, vertical_flip, horizontal_flip, addChannels]\n        self.data_type = data_type\n        self.audio_data = audio_data\n        self.fmin = fmin\n        self.fmax = fmax\n        self.length = length\n\n        for i in range(0, len(X)):\n            recording_id = X[i]\n            label = y[i]\n            mel_spec = self.audio_data[recording_id]\n            self.data.append(mel_spec)\n            self.labels.append(label)\n\n    def __len__(self):\n        return len(self.data)\n\n    def __getitem__(self, idx):\n        if self.data_type == \"train\":\n            aug = random.choice(self.augs)\n            data = aug(self.data[idx])\n        else:\n            data = addChannels(self.data[idx])\n        return data, self.labels[idx]\n\ndef horizontal_flip(img):\n    horizontal_flip_img = img[:, ::-1]\n    return addChannels(horizontal_flip_img)\n\ndef vertical_flip(img):\n    vertical_flip_img = img[::-1, :]\n    return addChannels(vertical_flip_img)\n\ndef addNoisy(img):\n    noise_img = util.random_noise(img)\n    return addChannels(noise_img)\n\ndef contrast_stretching(img):\n    contrast_img = exposure.rescale_intensity(img)\n    return addChannels(contrast_img)\n\ndef randomGaussian(img):\n    gaussian_img = gaussian(img)\n    return addChannels(gaussian_img)\n\ndef grayScale(img):\n    gray_img = rgb2gray(img)\n    return addChannels(gray_img)\n\ndef randomGamma(img):\n    img_gamma = exposure.adjust_gamma(img)\n    return addChannels(img_gamma)\n\ndef addChannels(img):\n    return np.stack((img, img, img))\n\ndef spec_to_image(spec):\n    spec = resize(spec, (224, 400))\n    eps = 1e-6\n    mean = spec.mean()\n    std = spec.std()\n    spec_norm = (spec - mean) / (std + eps)\n    spec_min, spec_max = spec_norm.min(), spec_norm.max()\n    spec_scaled = 255 * (spec_norm - spec_min) / (spec_max - spec_min)\n    spec_scaled = spec_scaled.astype(np.uint8)\n    spec_scaled = np.asarray(spec_scaled)\n    return spec_scaled\n\ndef get_model(num_labels):\n    model = EfficientNet.from_pretrained('efficientnet-b0', num_classes=num_labels)\n    model = model.to(device)\n    return model\n\n\ndef generate_submission():\n    members = []\n    for i in range(1, nfold):\n        member_model = get_model(num_labels)\n        member_model.load_state_dict(torch.load('./model'+str(i)+'.pt'))\n        member_model.eval()\n        members.append(member_model)\n\n    with open('submission.csv', 'w', newline='') as csvfile:\n        submission_writer = csv.writer(csvfile, delimiter=',')\n        submission_writer.writerow(['recording_id','s0','s1','s2','s3','s4','s5','s6','s7','s8','s9','s10','s11',\n                                   's12','s13','s14','s15','s16','s17','s18','s19','s20','s21','s22','s23'])\n\n        test_files = os.listdir('../input/rfcx-species-audio-detection/test/')\n\n        for i in tqdm(list(range(0, len(test_files)))):\n            data = load_test_file(test_files[i])\n            data = torch.tensor(data)\n            data = data.float()\n            if torch.cuda.is_available():\n                data = data.cuda()\n\n            output_list = []\n            for m in members:\n                output = m(data)\n                maxed_output = torch.max(output, dim=0)[0]\n                maxed_output = maxed_output.cpu().detach()\n                output_list.append(maxed_output)\n            avg_maxed_output = torch.mean(torch.stack(output_list), dim=0)\n\n            file_id = str.split(test_files[i], '.')[0]\n            write_array = [file_id]\n\n            for out in avg_maxed_output:\n                write_array.append(out.item())\n\n            submission_writer.writerow(write_array)\n","metadata":{"trusted":true},"outputs":[],"execution_count":null},{"cell_type":"code","source":"num_labels = 24\ndevice = 'cuda:0' if torch.cuda.is_available() else 'cpu'\nwarnings.filterwarnings('ignore')\nlearning_rate = 2e-4\nepochs = 20\nloss_fn = nn.CrossEntropyLoss()\nnfold = 5\nsr = 48000\nlength = 10 * sr\nfmin = 24000\nfmax = 0","metadata":{"trusted":true},"outputs":[],"execution_count":null},{"cell_type":"code","source":"data = pd.read_csv(\"../input/rfcx-species-audio-detection/train_tp.csv\")\n\nfmin = 24000\nfmax = 0\nfor i in range(0, len(data)):\n    if fmin > float(data.iloc[i]['f_min']):\n        fmin = float(data.iloc[i]['f_min'])\n    if fmax < float(data.iloc[i]['f_max']):\n        fmax = float(data.iloc[i]['f_max'])\n        \nfmin = int(fmin * 0.9)\nfmax = int(fmax * 1.1)\n\nlabel_list = []\ndata_list = []\naudio_data = {}\nfor i in tqdm(list(range(0, len(data)))):\n    recording_id = data.recording_id.values[i]\n    species_id = int(data.species_id.values[i])\n    data_list.append(recording_id)\n    label_list.append(species_id)\n\n    wav, sr = librosa.load('../input/rfcx-species-audio-detection/train/' + recording_id + '.flac', sr=None)\n    t_min = float(data.t_min.values[i]) * sr\n    t_max = float(data.t_max.values[i]) * sr\n    center = np.round((t_min + t_max) / 2)\n    beginning = center - length / 2\n    if beginning < 0:\n        beginning = 0\n    ending = beginning + length\n    if ending > len(wav):\n        ending = len(wav)\n        beginning = ending - length\n    slice = wav[int(beginning):int(ending)]\n    \n    spec=librosa.feature.melspectrogram(y=slice, sr=sr, fmin=fmin, fmax=fmax)\n    spec_db=librosa.power_to_db(spec, top_db=80)\n    \n    img = spec_to_image(spec_db)\n    \n    audio_data[recording_id] = img","metadata":{"trusted":true},"outputs":[],"execution_count":null},{"cell_type":"code","source":"def train(model, loss_fn, train_loader, valid_loader, epochs, optimizer, scheduler):\n    best_model_wts = copy.deepcopy(model.state_dict())\n    best_acc = 0.0\n    train_losses = []\n    valid_losses = []\n\n    postfix_label = \"\"\n    \n    for epoch in tqdm(range(1,epochs+1), postfix=postfix_label):\n        model.train()\n        batch_losses=[]\n        for _, data in enumerate(train_loader):\n            x, y = data\n            optimizer.zero_grad()\n            x = x.to(device, dtype=torch.float32)\n            y = y.to(device, dtype=torch.long)\n            y_hat = model(x)\n            loss = loss_fn(y_hat, y)\n            loss.backward()\n            batch_losses.append(loss.item())\n            optimizer.step()\n        train_losses.append(batch_losses)\n\n        model.eval()\n        batch_losses=[]\n        trace_y = []\n        trace_yhat = []\n        \n        for _, data in enumerate(valid_loader):\n            x, y = data\n            x = x.to(device, dtype=torch.float32)\n            y = y.to(device, dtype=torch.long)\n            y_hat = model(x)\n            loss = loss_fn(y_hat, y)\n            trace_y.append(y.cpu().detach().numpy())\n            trace_yhat.append(y_hat.cpu().detach().numpy())      \n            batch_losses.append(loss.item())\n        valid_losses.append(batch_losses)\n        trace_y = np.concatenate(trace_y)\n        trace_yhat = np.concatenate(trace_yhat)\n        accuracy = np.mean(trace_yhat.argmax(axis=1)==trace_y)\n\n        scheduler.step(np.mean(valid_losses[-1]))\n        if accuracy > best_acc:\n            best_acc = accuracy\n            best_model_wts = copy.deepcopy(model.state_dict())\n\n        postfix_label = \"epoch = %d, train_loss = %.5f, val_loss = %.5f, val_accuracy = %.5f\" % (epoch, np.mean(train_losses[-1]), np.mean(valid_losses[-1]), accuracy)\n\n    model.load_state_dict(best_model_wts)\n    return model","metadata":{"trusted":true},"outputs":[],"execution_count":null},{"cell_type":"code","source":"skf = KFold(n_splits=nfold, shuffle=True, random_state=32)\n\nfor fold_id, (train_index, val_index) in tqdm(list(enumerate(skf.split(data_list, label_list)))):\n    X_train = np.take(data_list, train_index)\n    y_train = np.take(label_list, train_index, axis = 0)\n    X_val = np.take(data_list, val_index)\n    y_val = np.take(label_list, val_index, axis = 0)\n\n    train_data = AudioData(X_train, y_train, \"train\", audio_data, fmin, fmax, length)\n    valid_data = AudioData(X_val, y_val, \"valid\", audio_data, fmin, fmax, length)\n    train_loader = DataLoader(train_data, batch_size=8, shuffle=True, drop_last=True)\n    valid_loader = DataLoader(valid_data, batch_size=8, shuffle=True, drop_last=True)\n\n    model = get_model(num_labels)\n    optimizer = torch.optim.Adam(model.parameters(), lr=learning_rate)\n    scheduler = torch.optim.lr_scheduler.ReduceLROnPlateau(optimizer, 'min', patience=3)\n    model = train(model, loss_fn, train_loader, valid_loader, epochs, optimizer, scheduler)\n    torch.save(model.state_dict(), \"./model\" + str(fold_id) + \".pt\")\n    \n    del train_data, valid_data, train_loader, valid_loader, model, X_train, X_val, y_train, y_val","metadata":{"trusted":true},"outputs":[],"execution_count":null},{"cell_type":"code","source":"def load_test_file(f):\n    wav, sr = librosa.load('../input/rfcx-species-audio-detection/test/' + f, sr=None)\n\n    segments = len(wav) / length\n    segments = int(np.ceil(segments))\n    \n    mel_array = []\n    \n    for i in range(0, segments):\n        if (i + 1) * length > len(wav):\n            slice = wav[len(wav) - length:len(wav)]\n        else:\n            slice = wav[i * length:(i + 1) * length]\n        \n        spec=librosa.feature.melspectrogram(y=slice, sr=sr, fmin=fmin, fmax=fmax)\n        spec_db=librosa.power_to_db(spec,top_db=80)\n\n        img = spec_to_image(spec_db)\n        mel_spec = np.stack((img, img, img))\n        mel_array.append(mel_spec)\n    \n    return mel_array","metadata":{"trusted":true},"outputs":[],"execution_count":null},{"cell_type":"code","source":"generate_submission()","metadata":{"trusted":true},"outputs":[],"execution_count":null}]}